Gartner now pegs the enterprise AI coding agents market at somewhere between $9.8 billion and $11 billion in annualized value, according to research published in Gartner’s May 2026 report. That number alone is not the interesting part. What matters more for anyone buying these tools right now is who is selling them, and that list has quietly changed. The vendors building the frontier models — Anthropic, OpenAI, Google — are no longer content to license their models to coding-tool startups. They are shipping competing products directly, and in at least one case, out-earning the startups they used to supply.
Sizing the enterprise AI coding agent market
Depending on which firm you ask, you get a different number. Gartner’s $9.8-11 billion figure counts enterprise coding agents narrowly. Mordor Intelligence, using a wider definition that includes adjacent code tooling, puts 2026 at $14.18 billion, growing at a 26.44% CAGR toward $45.83 billion by 2031. Both firms agree on the underlying signal: two-thirds of technology leaders surveyed by Mordor said AI-generated or substantially refactored code made up 51-75% of their team’s weekly output in 2026. This is not a pilot-stage technology anymore. It is production infrastructure, and the money is following accordingly.
How frontier labs became coding agent vendors
Anthropic launched Claude Code in May 2025. By November it had crossed $1 billion in annualized revenue. By February 2026 that figure had reached $2.5 billion — over half of Anthropic’s entire enterprise revenue, according to reporting from AgentMarketCap. OpenAI answered with a relaunched Codex, which hit 2 million weekly active users by March 2026, and then bought Astral — the team behind the Python tools uv, Ruff and ty — folding it directly into Codex’s toolchain. GitHub Copilot, the incumbent, has stalled at 4.7 million paid subscribers despite 76% developer awareness of the product. Awareness clearly isn’t the bottleneck anymore.
The pattern is straightforward once you see it. A company that trains the underlying model gets first access to every capability improvement, watches how agents actually behave against real codebases before anyone else does, and can fold safety work into the model rather than bolting it onto an application built on top of someone else’s API. Selling that advantage to a third-party coding-tool vendor was never going to last once the labs decided the application layer was worth owning themselves.
Cursor’s response: stop renting the model
Cursor reached a $50 billion valuation and $2 billion in annualized revenue before Claude Code’s launch reshaped the field around it. Its response has been to start training its own frontier model, reportedly on Nvidia’s Colossus cluster. That is a strange move for a company that built its business on giving developers a choice of OpenAI, Anthropic, Gemini or xAI models inside one interface — and it only makes sense as a hedge against exactly the dynamic described above. If the model provider can always undercut you on integration depth, the only way to compete on even footing is to become a model provider too.
The real fight is over orchestration, not autocomplete
Six months ago, most developers worked with a single assistant in a tight loop: prompt, wait, review, repeat. That loop is disappearing. Databricks research cited by Reinventing.ai found multi-agent workflow deployments grew more than 300% in recent months as companies moved from pilots into production. Gartner expects more than 65% of engineering teams using agentic coding to treat the IDE itself as optional by 2027, as agents take over planning, review, testing and deployment rather than just code completion.
That shift is also why OpenAI, Anthropic and Google jointly founded the Agentic AI Foundation under the Linux Foundation in February 2026, bringing the Model Context Protocol, the goose framework and the AGENTS.md specification under one roof. Fierce competitors rarely standardize infrastructure together unless they all expect the market to fragment without it.
What engineering leaders should actually weigh
Growth numbers this large tend to hide the failure rate sitting underneath them. Gartner also forecasts that over 40% of agentic AI projects will be canceled by 2027 due to cost, unclear value or governance gaps. Before signing another annual contract, a few questions are worth answering directly:
- Does this vendor control its own model, or is it exposed if a frontier lab decides to compete with it directly?
- How much of the value is in code generation versus orchestration across planning, review and testing — and is that gap closing or widening?
- What happens to your workflows and integrations if the tool is acquired, discontinued or repriced, given how quickly this list of vendors has already reshuffled?
- Is adoption tracked against a real productivity baseline, or is “AI-generated code” being counted as progress on its own?
None of this argues against adopting these tools. The productivity case is real, and the vendors on both sides of the model-versus-application divide are shipping genuinely useful products. But the ground under this market is moving fast enough that the vendor you standardize on this quarter may be a very different company, selling a very different product, by the time your contract renews. Evaluate the roadmap as carefully as the demo.